AI and Autonomous Systems for Assessing Biodiversity and Ecosystem Health

Global biodiversity decline is a critical issue facing humanity. Monitoring biodiversity is essential for understanding the impact of this decline on the ecosystem services we depend on, as well as the impact of human activities and climate change on biodiversity.

Project status

Ongoing

Introduction

Biodiversity is in decline globally, with one in six species currently facing extinction in the UK. Monitoring biodiversity is therefore essential for understanding the impact of this decline on the ecosystem services we depend on, as well as the impact of human activities and climate change on biodiversity.  

Core to both conservation research and action is the monitoring of species and ecosystem health, helping us understand how human activities and interventions affect biodiversity. Traditionally, this has been done by manual sampling and interpretation by experts, yielding high-quality, but resource-expensive, biodiversity data. Therefore, there is widespread anticipation for an AI-led revolution in biodiversity monitoring to scale up monitoring efforts and to help meet national and international biodiversity targets.

At the Turing, we are developing new AI methods together with our partners to monitor biodiversity at scale. For this project, the focus is on monitoring challenges that can be catalysed by AI-powered upscaling: real-time monitoring of marine plankton, predicting habitat condition from Earth Observation data and adaptive sampling by Autonomous Underwater Vehicles (AUVs). Crucial to the project is the collaboration with domain experts to provide accurately annotated data for training machine learning models, and to align our developments with the priorities from government, industry, and academia.

Explaining the science

We will leverage connections between partners and key organisations to promote a Turing-led vision for AI-driven biodiversity monitoring , through existing events, new activities  and (inter)national initiatives, engaging with industry (instrumentation / software developers, water companies), government (Defra and Environment Agency as national monitoring bodies), research (HEIs, funders, data centres), and 3rd sector (citizen science networks). 
 
Addressing outstanding technical challenges will demonstrate how AI research can deliver solutions for biodiversity monitoring:  

Enhancing real-time plankton monitoring

Plankton plays a critical role in indicating the health of aquatic ecosystems due to its sensitivity to environmental changes and short life cycles. The advent of AI technologies and high-speed imaging instruments offers a promising avenue to enhance the accuracy and efficiency of plankton monitoring. We are utilising data from the  Plankton Imaging (Pi-10) instrument deployed on the  RV Cefas Endeavour to develop a near real-time monitoring application using SOTA AI techniques. 

Building on the work of the Turing Data Study Group (DSG), we aim to improve the DSG classifier for real-time access to plankton imager data (RAPID), a suite of tools designed for use with the Plankton Analytics Pi-10 instrument. The first version of the RAPID Plankton software was operational within six months and deployed on the RV Cefas Endeavour. This open-source software can be adapted to classify images of various marine objects and species, revolutionizing how scientists study the oceans. Further developments in this project include the deployment of a dashboard, which provides real-time plankton metrics. Additionally, we are developing explainable AI (xAI) interpretations to collaborate with ecologists, helping them understand how AI works and ensuring the AI models are transparent and interpretable. 

Predicting habitat condition from satellite imagery

We are developing new AI methods that can meet the demand for resource-efficient terrestrial biodiversity monitoring, by integrating two data sources available at scale: remote sensing and citizen-science data. Remote sensing data, such as satellite imagery, have wide coverage and are routinely used for mapping habitats. At the same time, wildlife records collected on the ground by volunteers give high-quality, but local and resource-intensive, measurements of habitat quality. For example, some butterflies are only found in healthy grasslands, and are therefore good indicators of ecosystem health.  

As part of this research, we aim to bridge the scale from local recordings to global satellite coverage, by developing a predictive model of species biodiversity based on satellite data. This model can be used for locating high-biodiversity habitats, at high spatial and temporal resolution in a standardised way. This is important as more nature needs to be protected to reach the UN’s goal of protecting 30% of land and sea by 2030. 

 Satellite image analysis in areas where wildlife sightings were recorded by volunteers. Satellite images from ESA Sentinel-2, animal icons by Ian Rahmadi Kurniawan (Noun project).

Adaptive sampling

Autonomous Underwater Vehicles (AUVs) currently rely on ‘human pilots’ and follow predetermined waypoints. This research aims to equip pilot-less AUVs with advanced information-gathering capabilities to optimise the quantity/quality of the data collected, while satisfying safety and environmental constraints using newly developed algorithms (greedy approaches, AI planning, and reinforcement learning). 

Rethinking evaluation of AI performance for biodiversity

We are developing new computer vision methods to analyse images from satellites and species observations, in conjunction with Turing's Scivision to enable the broader research community to use these tools more widely. Furthermore, we are rethinking traditional computer vision metrics for the evaluation of real and synthetic biodiversity data to account for the structure of natural communities. With this, we seek to provide guidance to the scientific community on best practices for evaluating AI biodiversity monitoring.  

Project aims

  • Engage with the UK’s principal national biodiversity research and monitoring organisations to identify key scaling challenges that AI could address. 

  • Develop new computer vision techniques that can accurately and efficiently perform tasks necessary for biodiversity monitoring. 

  • Curate high-quality biodiversity data sets for the continual development of AI methods by the research community. 

  • Deploy new AI methods together with partner organisations to pioneer real-world AI biodiversity monitoring applications, demonstrating the applicability and potential of AI to scale up biodiversity monitoring.

Applications

This project will significantly enhance Turing activity in biodiversity research. It will strengthen links across NERC centres, increasing AI and data science research in these organisations that undertake and coordinate significant environmental research and data stewardship.  

Through its variety of real-world deployments, this project will demonstrate how Turing AI research can advance biodiversity monitoring, necessary for tracking our progress to achieve (inter)national biodiversity targets.  

Organisers

Matt Fry

Environmental Informatics Manager, UK Centre for Ecology and Hydrology

Harry Meacher

Programme Manager – Environment & Sustainability Grand Challenge

Researchers and collaborators

Matt Fry

Environmental Informatics Manager, UK Centre for Ecology and Hydrology

Contact info

Monica Vakil-Dewar
[email protected]

Participating researchers

  • Sophie Pitois, CEFAS
  • Robert Blackwell, CEFAS
  • James Scott, CEFAS
  • Sara Bernardini, RHUL
  • Steven Thackeray, UKCEH
  • Amulya Chevuturi, UKCEH
  • Michael Pocock, UKCEH
  • Simon Rolph, UKCEH
  • Diana Bowler, UKCEH
  • Blair Thornton, University of Southampton
  • Thijs van der Plas, Turing
  • Noushin Eftekhari, Turing
  • Jen Durden, NOC
  • Veerle Huvenne, NOC
  • Sari Giering, NOC
  • Catherine Wardell, NOC
  • Mojtaba Masoudi, NOC
  • Zonghua Liu, NOC

Collaborating organisations

  • CEFAS
  • UKCEH
  • NOC